Model individualized for real-time operator functional state assessment

• Published in: Advances in Human Aspects of Aviation pp. 431 - 440
• Format: Book Chapter
• Language: English
• Published: CRC Press 2013
• Subjects: Generation Air Transportation System; TRACON; Dataset; FAA; ASRS Report; Crew Resource Management; Convective Weather; MIT Lincoln Lab; Air Route Traffic Control Center; Disengaged; ASRS; Validation Dataset; CRM Training; Federal Aviation Administration; NASA Ames Research Center; Low Workload Conditions; Situation Awareness; Alert; Air Traffic Management System; Model Individualization; Model Development Support; OFS; Flight Deck Design; CRM
• DOI: 10.1201/b12321-48

Proper assessment of Operator Functional State (OFS) and appropriate workload modulation offer the potential to improve mission effectiveness and aviation safety in both overload and under-load conditions. Although a wide range of research has been devoted to building OFS assessment models, most of the models are based on group statistics and little or no research has been directed towards model individualization, i.e., tuning the group statistics based model for individual pilots. Moreover, little emphasis has been placed on monitoring whether the pilot is disengaged during low workload conditions. The primary focus of this research is to provide a real-time engagement assessment technique considering individual variations in an aviation environment. This technique is based on an advancedmachine learning technique, called enhanced committee machine. We have investigated two different model individualization approaches: similarity-based and dynamic ensemble selection-based. The basic idea of the similarity-based technique is to find similar subjects from the training data pool and use their data together with the limited training data from the test subject to build an individualized OFS assessment model. The dynamic ensemble selection dynamically select data points in a validation dataset (with labels) that are adjacent to each test sample, and evaluate all the trained models using the identified data points. The best performing models will be selected and maximum voting can be applied to perform individualized assessment for the test sample. To evaluate the developed approaches, we have collected data from a high fidelity Boeing 737 simulator. The results show that the performance of the dynamic ensemble selection approach is comparable to that achieved from an individual model (assuming sufficient data is available from each individual).